Descriptions:
Veteran AI walks through two officially released LTX 2.3 LoRAs — the Union Control LoRA and the Motion Track LoRA — demonstrating how they enable precise video-to-video generation inside ComfyUI. The Union Control LoRA consolidates pose maps, depth maps, and edge maps into a single model with a default reference strength of 0.5, while the Motion Track LoRA allows creators to define subject and camera movement by drawing trajectory curves directly, making zoom-in, zoom-out, and scene transition effects straightforward to execute.
The tutorial covers the full technical pipeline in detail: reference video should be scaled to a 544px shortest side, control images generated at 512px, and reference images scaled to a 1536px longest edge with a compression factor of 18. A key practical tip is substituting Kijai’s LTX model implementation for the native ComfyUI loader — the swap reduces VRAM consumption without a detectable quality penalty. The creator also flags that setting LoRA weight to 0.5 (rather than the default 1.0) is essential when working with high-motion source footage, as full strength causes character collapse.
The most technically distinctive segment covers LTX 2.3’s built-in prompt generator: a ComfyUI node called “Text Generate LTX To Prompt” that leverages CLIP — itself a large language model — to expand structured prompt engineering templates into detailed, cinematically precise generation prompts. Rather than feeding short freeform descriptions, the recommended workflow passes a full prompt engineering template through the node alongside a reference image, producing prompts with significantly richer narrative and motion specificity than hand-written alternatives. All workflows are available on RunningHub’s cloud ComfyUI platform.
📺 Source: Veteran AI · Published March 18, 2026
🏷️ Format: Tutorial Demo
